--- title: "You Don't Need a Powerful PC to Run AI. You Need a $20 Server." description: "Running AI on your own laptop chokes it, because local models are heavy and a good agent runs for hours. A 12-year automation consultant on why the smarter move is a cheap cloud server you drive from your phone, what it really costs, and how to set it up." url: https://techwalk.blog/Posts/run-ai-on-a-cheap-vps-july-2026/ content_type: text/markdown html_version: https://techwalk.blog/Posts/run-ai-on-a-cheap-vps-july-2026/ --- > This is the Markdown edition of this page, served to AI agents and for quick reading. The full HTML version, with images, interactive data charts and more, is at https://techwalk.blog/Posts/run-ai-on-a-cheap-vps-july-2026/ ![A light laptop and a phone connecting to a small, calm server in the cloud that does all the heavy AI work](/Posts/run-ai-on-a-cheap-vps-july-2026/main_image.png) # You Don't Need a Powerful PC to Run AI. You Need a $20 Server. Most people run AI on the same laptop they work on, then wonder why everything grinds to a halt. You do not need a powerful machine. You need a cheap server and a browser. Here is the setup. Published: July 2026 • \~7-minute read Open a serious AI model on your laptop and you can feel it happen. The fans spin up. The cursor starts to stutter. The thing you were actually doing, the video export, the twelve browser tabs, the design file, slows to a crawl. I know people who keep a second laptop running just so they can keep working while the AI chews through the first one. I do this for a living, and for a long time I did it the same way: everything on one machine, all at once. It is the obvious way to start. It is also the reason your computer feels slow the moment the work gets interesting. The good news is that the fix is not a more expensive laptop. It is the opposite. ## Your laptop is doing the one job it should not Here is the thing nobody says out loud when they tell you to “run AI locally.” A capable model is genuinely heavy. To run an open model like Llama 3.1 on your own machine at a usable quality, you are looking at around 16 GB of memory for the small 8B version, and something between 40 and 140 GB for the 70B. That is not a laptop. That is a workstation, and a loud one. To run an open model on your own computer you need real memory: about 16 GB for Llama 3.1 8B, and roughly 40 GB at 4-bit (up to \~140 GB at full precision) for 70B. A cloud CLI client like Claude Code or Codex only talks to a hosted model, so it holds under a gigabyte; the heavy compute runs on the server. Sources: [Hugging Face](https://huggingface.co/blog/llama31), [Ollama](https://ollama.com/library/llama3.1) (2024-2026). Now look at the other bar. A command line tool like **Claude Code** or **Codex** does not hold the model at all. It just talks to a hosted one over the internet, so it sits under a gigabyte while the heavy lifting happens on someone else’s hardware. That single gap is the whole idea of this article. The expensive part of AI does not have to live on your computer. ## The reframe: the AI does not have to run on your machine Once you accept that, the problem changes shape. You are not trying to cram a workstation’s worth of compute into a thin laptop. You are trying to put the heavy work somewhere it belongs, and then reach it from wherever you are. That “somewhere” is a small server in the cloud. You rent it by the month, it runs around the clock, and it never has to share a processor with your actual work. Your laptop goes back to being a screen and a keyboard, which is all it needed to be. ## A cheap server does far more than you would guess This is the part that surprises people. You do not need anything powerful. I run my own setup on a **netcup** VPS. Even when it is rendering video, which is about the heaviest thing I throw at it, the memory sits under 5 GB. The processor only really wakes up during a render. Most of the time it is coasting, running many sessions at once without complaint. The only skill you actually need is a good session manager, something like tmux, so your work keeps running when you close the lid. And the prices are almost silly. Here is netcup’s current ladder. What netcup's VPS plans give you each month, gross (incl. 19% VAT). The 4-vCPU, 8 GB VPS 1000 at about nine euros is already plenty for many AI sessions; the 8-vCPU, 16 GB VPS 2000 (highlighted) is the tier we run. Source: [netcup VPS pricing](https://www.netcup.com/en/server/vps) (July 2026). The plan I would start most people on is the VPS 1000: four cores and 8 GB of memory for about nine euros a month. That is already more than enough to run several AI sessions with room to spare. The tier I personally run, the VPS 2000, doubles that for around sixteen. Either one costs less than a couple of coffees and does more work than the laptop you are reading this on. You do not have to use netcup. Hetzner, Hostinger and DigitalOcean all land in the same ballpark. The point is the category, not the logo: a $10 to $20 server is plenty. ## How it actually works The setup is less involved than people expect. * **Rent the server.** Pick a cheap Linux VPS. A few euros a month is fine to begin. * **Install one command line tool.** Either Claude Code or Codex, whichever you prefer, installed directly on the server against your own account. * **Connect from anywhere.** Open a session over SSH or a browser, start it inside tmux, and it keeps running whether you are at your desk or on your phone on a train. That is genuinely it. Your editing, your automations, your long agent runs, all of it now lives on a machine that exists to do exactly that, at any hour, without touching the computer you actually work on. ## What you keep on your side: a cheap, light laptop Here is the money version of the argument. If you wanted to run this AI work locally, you would need a laptop with enough memory to hold a model. A 48 GB MacBook Pro is roughly $2,899\. That is the price of admission for doing it the heavy way. A laptop with enough memory to run AI locally, a 48 GB MacBook Pro, is about $2,899 up front. Two years of a netcup VPS 1000 is roughly $230 (about $430 for the VPS 2000), it runs around the clock, and your own laptop stays cheap and light. Sources: [Micro Center](https://www.microcenter.com/product/688643/apple-macbook-pro-16-z1ft000ce-%28late-2024%29-162-laptop-computer-space-black), [netcup](https://www.netcup.com/en/server/vps) (July 2026). Two years of the cheap server is a rounding error next to that, and the server is always on and running many sessions while it does it. So you buy the light, cheap laptop you actually enjoy carrying, and you let a $10 box in a data center be the powerful one. You get the best of both, and you stop paying the “powerful laptop” tax to do work your laptop was never the right place for. ## The one honest exception I will be straight about the one thing that does not fit neatly on a server: tasks that need a real, visible browser or a specific desktop app. Some automations genuinely want a screen. Even that has an answer now. You can drive a cloud browser, a real browser running in the cloud, over the internet, so the fingerprinted, logged-in browsing still happens off your machine. Unless you specifically need a native desktop application in front of you, you do not need a big computer sitting on your desk at all. ## Build it once, then redeploy it anywhere There is a nice trick that makes this close to zero effort the second time. Once you have a server set up the way you like, with your tools installed and configured, you save it as a snapshot, an image of the whole machine. After that, spinning up a new one is a click. You can even build the image on a tiny server rented by the hour, save the snapshot, and destroy the original, so you pay only for the minutes it took to set up. Your setup becomes portable: redeploy it on a bigger box when you need more power, or in a different region when you need to be closer to something, and carry on exactly where you left off. ## Do it with the doors locked One warning, because it matters. The reason a separate server is smart is not only performance, it is safety. An AI agent with real access can do real damage, and you do not want it running loose on the machine that holds your passwords and your bank access. I wrote about that in detail in [The “Accept All” Trap](/Posts/agentic-ai-own-server-june-2026/), including a client who ran up a five-figure API bill from a single leaked key. The short version: give the AI its own room. Use SSH keys, not passwords. Set hard spending limits and billing alerts on your API keys before you need them. Keep that server holding nothing that would hurt to lose. A cheap server is not just cheaper than a powerful laptop, it is a safer place to let an agent work. ## The bottom line The instinct is to buy your way out of the problem with a more powerful computer. It is the wrong instinct. The AI does not need to run on your machine, and the moment you move it off, your laptop stops choking, your work stops stalling, and the whole thing gets cheaper. Rent the cheap server. Put the heavy work there. Keep the light laptop you like. That is the entire move, and it costs about the price of lunch. ## 🌍 Join our free IT digital nomad community I share the exact setups I use to run AI on cheap servers, find clients, and deliver real work from anywhere. Come see how it actually works, and grab the templates. [Join the FREE community](https://link.brevibot.com/techwalk-community) --- ## About the author [Sergiu Butnaru](https://brevibot.com/sergiu) has spent 12 years building automation and AI systems for businesses and co-founded Brevibot. He writes TechWalk on AI, automation, security and the digital-nomad lifestyle. Connect: [Facebook](https://www.facebook.com/sergiu.ser.butnaru) | [LinkedIn](https://www.linkedin.com/in/sergiu-butnaru-brevibot) | [Instagram](https://www.instagram.com/sergiu.butnaru.it). ## Join the community Join the free IT Digital Nomad community to see exactly how we use these AI systems to find clients, deliver real work, and build income with AI right now: https://link.brevibot.com/techwalk-community --- Published by [Brevibot](https://brevibot.com). TechWalk: https://techwalk.blog